Ten Companies, No Hospitals

Ten companies sat down with federal regulators in July to demonstrate clinical AI. Not one of them runs a hospital.

That is the piece. If you chair a governance committee, do not spend your energy objecting to the meeting. Spend it noticing that the evidence standard taking shape in that room will reach your organization as a finished product with a clearance letter attached, and get your position on the record while there is still a record being built.

What happened

On July 8, officials from the Food and Drug Administration and the Centers for Medicare and Medicaid Services hosted leaders from ten companies at the FDA's White Oak headquarters for what the agenda called a clinical AI demo day. The meeting was not publicly announced. STAT's Mario Aguilar reported it on August 5, working from an agenda reviewed by the publication.

• Anthropic
• Counsel Health
• Curai
• K Health
• Microsoft AI
• Amazon One Medical
• Doctronic
• Ellipsis Health
• Hippocratic AI
• Welldoc

Several are backed by Andreessen Horowitz or Khosla Ventures. Four of them, Counsel Health, K Health, Hippocratic AI, and Ellipsis Health, later described to STAT what they had told regulators.

Let me be precise about what I am not arguing, because this is where these conversations usually go sideways. Vendors belong in front of regulators. A regulator who has never seen the product cannot write a rule that survives contact with the product, and the alternative to meetings like this one is not a better process. It is a slower and more ignorant one. The demo day was not improper. The agencies are also running a parallel track: on July 23, the White House Office of Science and Technology Policy, the FDA, and ONC convened outside experts for a one-month sprint toward a consensus set of principles on clinical AI benchmarking. So this is not a story about regulatory capture, and anyone selling it to you that way is overreaching.

The guest list is the finding

Every organization on that agenda builds or sells technology.

No hospital appears on it. No health system. No physician membership organization, no nursing organization, no state licensing board, no patient safety organization.

Here is why the composition matters more than the meeting. A demo answers the question "does it work." Deployment answers a different question, which is "does it work here, on this staffing model, in this workflow, at this hour, with these people carrying the liability when it does not." Those questions have different evidence standards, and only one of them gets asked in a room where every participant is showing you their product working.

I practice internal medicine. I trained in surgery. The gap between a tool performing in a demonstration and a tool performing on a floor at 2:00 AM with two nurses covering thirty beds is enormous, and almost nothing about it is technical. It is operational. The people who could have described it were not on the list.

Before you write to me about Amazon One Medical

Fair. One Medical employs practicing clinicians, and if you want to argue that clinical practice had a seat, that is the name you would reach for. Here is why I do not think it closes the gap.

One Medical is a primary care organization owned by a technology company, running a delivery model it designed itself. It holds its own liability, inside a system it controls end to end, with a patient population it selected. That is a real and useful perspective. It is not the perspective of a 400-bed community hospital that will be handed a cleared tool, a vendor contract it did not negotiate, and a nursing staff who had no part in choosing any of it.

Both perspectives should be in the room. Only one was.

Four things to do before your next quarterly meeting

A governance checklist your committee can run in 60 minutes

The Utah footnote, corrected

Doctronic was on that guest list, and it is the name most people in my feed recognize, usually inaccurately. Let me set the record straight, because I have watched this one get repeated wrong for months and the wrong version is doing real work in people's arguments.

Utah's Office of Artificial Intelligence Policy entered a twelve-month agreement with Doctronic, announced in January 2026. It is a prescription renewal pilot. The AI cannot write a new prescription. It cannot touch a controlled substance. It cannot change a dose, a frequency, or a treatment plan. The formulary runs to roughly 190 to 200 non-controlled drugs for chronic conditions.

And it is still in Phase 1, where a licensed practitioner authorizes every single request. The autonomous step, where a renewal routes straight to the pharmacist, is Phase 2. It requires an approval the state has not granted.

The Utah Medical Licensing Board learned about the agreement after it launched, wrote that proceeding without consulting the board "potentially places Utah citizens at risk," and recommended immediate suspension. The Department of Commerce declined. Six-month figures are public: the AI recommended approval and routed to a physician in 72% of cases, escalated on its own in 28%, and physicians agreed the escalation was warranted 69% of the time.

Read that last number again. Nearly a third of the time the system escalated, the reviewing physician thought it did not need to. Call that what it is: a calibration problem, not a scandal. Calibration problems are exactly what Phase 1 exists to surface and fix before Phase 2, which is the argument for seating the licensing board at the start rather than letting it read the press release.

So the autonomous AI physician that half of LinkedIn argues about does not exist yet. Not in Utah, not anywhere. What exists is a formulary of about 190 chronic-condition drugs, a queue, a calibration gap, and a licensing board that found out from a press release.

Sarah Matt, MD, MBA is a surgery-trained physician-executive currently practicing internal medicine (charity care). She advises health systems and health-tech companies on clinical AI governance and implementation through Vital Werks.